Nemotron 3.5 Lightning

NVIDIA Nemotron 3.5 Lightning is an open-weights large language model optimized for high-efficiency, long-running agentic workflows. It utilizes a Mixture-of-Experts (MoE) architecture with 30 billion total parameters and 3 billion active parameters per token. This design allows the model to balance complex reasoning capabilities with the low-latency performance required for high-volume task execution in multi-agent systems.

The model's architecture is a hybrid LatentMoE structure that integrates Mamba-2, standard MoE, and Attention layers. It incorporates Multi-Token Prediction (MTP), where additional layers are trained to predict multiple future tokens simultaneously. This feature provides richer training signals and facilitates efficient speculative decoding, significantly improving generation throughput compared to traditional dense models of similar size.

Optimized for agentic tasks, Nemotron 3.5 Lightning is specifically tuned for tool calling, output validation, result formatting, and sub-agent delegation. It supports a massive 1 million token context window, enabling deep reasoning across extensive documents and long-duration conversational histories. The model is trained on a vast corpus of over 20 trillion tokens, including support for English and 19 additional spoken languages, alongside 43 programming languages.

In addition to general-purpose language tasks, the model is evaluated for its proficiency in math, coding, and scientific reasoning. It has demonstrated notable improvements in agentic benchmarks over previous generations, particularly in complex software engineering and long-context retrieval tasks. Its hybrid architecture and open weights make it a versatile foundation for domain-specific fine-tuning in fields such as cybersecurity, finance, and healthcare.